2 * NTP client/server, based on OpenNTPD 3.9p1
4 * Author: Adam Tkac <vonsch@gmail.com>
6 * Licensed under GPLv2, see file LICENSE in this tarball for details.
8 * Parts of OpenNTPD clock syncronization code is replaced by
9 * code which is based on ntp-4.2.6, whuch carries the following
12 ***********************************************************************
14 * Copyright (c) University of Delaware 1992-2009 *
16 * Permission to use, copy, modify, and distribute this software and *
17 * its documentation for any purpose with or without fee is hereby *
18 * granted, provided that the above copyright notice appears in all *
19 * copies and that both the copyright notice and this permission *
20 * notice appear in supporting documentation, and that the name *
21 * University of Delaware not be used in advertising or publicity *
22 * pertaining to distribution of the software without specific, *
23 * written prior permission. The University of Delaware makes no *
24 * representations about the suitability this software for any *
25 * purpose. It is provided "as is" without express or implied *
28 ***********************************************************************
32 #include <netinet/ip.h> /* For IPTOS_LOWDELAY definition */
33 #include <sys/timex.h>
34 #ifndef IPTOS_LOWDELAY
35 # define IPTOS_LOWDELAY 0x10
38 # error "Sorry, your kernel has to support IP_PKTINFO"
42 /* Verbosity control (max level of -dddd options accepted).
43 * max 5 is very talkative (and bloated). 2 is non-bloated,
44 * production level setting.
49 #define RETRY_INTERVAL 5 /* on error, retry in N secs */
50 #define RESPONSE_INTERVAL 15 /* wait for reply up to N secs */
52 #define FREQ_TOLERANCE 0.000015 /* % frequency tolerance (15 PPM) */
54 #define MINPOLL 4 /* % minimum poll interval (6: 64 s) */
55 #define MAXPOLL 12 /* % maximum poll interval (12: 1.1h, 17: 36.4h) (was 17) */
56 #define MINDISP 0.01 /* % minimum dispersion (s) */
57 #define MAXDISP 16 /* maximum dispersion (s) */
58 #define MAXSTRAT 16 /* maximum stratum (infinity metric) */
59 #define MAXDIST 1 /* % distance threshold (s) */
60 #define MIN_SELECTED 1 /* % minimum intersection survivors */
61 #define MIN_CLUSTERED 3 /* % minimum cluster survivors */
63 #define MAXDRIFT 0.000500 /* frequency drift we can correct (500 PPM) */
65 /* Clock discipline parameters and constants */
66 #define STEP_THRESHOLD 0.128 /* step threshold (s) */
67 #define WATCH_THRESHOLD 150 /* stepout threshold (s). std ntpd uses 900 (11 mins (!)) */
68 /* NB: set WATCH_THRESHOLD to ~60 when debugging to save time) */
69 #define PANIC_THRESHOLD 1000 /* panic threshold (s) */
71 /* Poll-adjust threshold.
72 * When we see that offset is small enough compared to discipline jitter,
73 * we grow a counter: += MINPOLL. When it goes over POLLADJ_LIMIT,
74 * we poll_exp++. If offset isn't small, counter -= poll_exp*2,
75 * and when it goes below -POLLADJ_LIMIT, we poll_exp--
77 #define POLLADJ_LIMIT 30
78 /* If offset < POLLADJ_GATE * discipline_jitter, then we can increase
79 * poll interval (we think we can't improve timekeeping
80 * by staying at smaller poll).
82 #define POLLADJ_GATE 4
83 /* Compromise Allan intercept (s). doc uses 1500, std ntpd uses 512 */
87 /* FLL loop gain [why it depends on MAXPOLL??] */
88 #define FLL (MAXPOLL + 1)
89 /* Parameter averaging constant */
98 NTP_MSGSIZE_NOAUTH = 48,
99 NTP_MSGSIZE = (NTP_MSGSIZE_NOAUTH + 4 + NTP_DIGESTSIZE),
102 MODE_MASK = (7 << 0),
103 VERSION_MASK = (7 << 3),
107 /* Leap Second Codes (high order two bits of m_status) */
108 LI_NOWARNING = (0 << 6), /* no warning */
109 LI_PLUSSEC = (1 << 6), /* add a second (61 seconds) */
110 LI_MINUSSEC = (2 << 6), /* minus a second (59 seconds) */
111 LI_ALARM = (3 << 6), /* alarm condition */
114 MODE_RES0 = 0, /* reserved */
115 MODE_SYM_ACT = 1, /* symmetric active */
116 MODE_SYM_PAS = 2, /* symmetric passive */
117 MODE_CLIENT = 3, /* client */
118 MODE_SERVER = 4, /* server */
119 MODE_BROADCAST = 5, /* broadcast */
120 MODE_RES1 = 6, /* reserved for NTP control message */
121 MODE_RES2 = 7, /* reserved for private use */
124 //TODO: better base selection
125 #define OFFSET_1900_1970 2208988800UL /* 1970 - 1900 in seconds */
127 #define NUM_DATAPOINTS 8
140 uint8_t m_status; /* status of local clock and leap info */
142 uint8_t m_ppoll; /* poll value */
143 int8_t m_precision_exp;
144 s_fixedpt_t m_rootdelay;
145 s_fixedpt_t m_rootdisp;
147 l_fixedpt_t m_reftime;
148 l_fixedpt_t m_orgtime;
149 l_fixedpt_t m_rectime;
150 l_fixedpt_t m_xmttime;
152 uint8_t m_digest[NTP_DIGESTSIZE];
162 len_and_sockaddr *p_lsa;
164 /* when to send new query (if p_fd == -1)
165 * or when receive times out (if p_fd >= 0): */
168 uint32_t lastpkt_refid;
169 uint8_t lastpkt_status;
170 uint8_t lastpkt_stratum;
171 uint8_t reachable_bits;
172 double next_action_time;
174 double lastpkt_recv_time;
175 double lastpkt_delay;
176 double lastpkt_rootdelay;
177 double lastpkt_rootdisp;
178 /* produced by filter algorithm: */
179 double filter_offset;
180 double filter_dispersion;
181 double filter_jitter;
182 datapoint_t filter_datapoint[NUM_DATAPOINTS];
183 /* last sent packet: */
193 /* Insert new options above this line. */
194 /* Non-compat options: */
198 OPT_l = (1 << 7) * ENABLE_FEATURE_NTPD_SERVER,
203 /* total round trip delay to currently selected reference clock */
205 /* reference timestamp: time when the system clock was last set or corrected */
207 /* total dispersion to currently selected reference clock */
210 double last_script_run;
213 #if ENABLE_FEATURE_NTPD_SERVER
218 /* refid: 32-bit code identifying the particular server or reference clock
219 * in stratum 0 packets this is a four-character ASCII string,
220 * called the kiss code, used for debugging and monitoring
221 * in stratum 1 packets this is a four-character ASCII string
222 * assigned to the reference clock by IANA. Example: "GPS "
223 * in stratum 2+ packets, it's IPv4 address or 4 first bytes of MD5 hash of IPv6
227 /* precision is defined as the larger of the resolution and time to
228 * read the clock, in log2 units. For instance, the precision of a
229 * mains-frequency clock incrementing at 60 Hz is 16 ms, even when the
230 * system clock hardware representation is to the nanosecond.
232 * Delays, jitters of various kinds are clamper down to precision.
234 * If precision_sec is too large, discipline_jitter gets clamped to it
235 * and if offset is much smaller than discipline_jitter, poll interval
236 * grows even though we really can benefit from staying at smaller one,
237 * collecting non-lagged datapoits and correcting the offset.
238 * (Lagged datapoits exist when poll_exp is large but we still have
239 * systematic offset error - the time distance between datapoints
240 * is significat and older datapoints have smaller offsets.
241 * This makes our offset estimation a bit smaller than reality)
242 * Due to this effect, setting G_precision_sec close to
243 * STEP_THRESHOLD isn't such a good idea - offsets may grow
244 * too big and we will step. I observed it with -6.
246 * OTOH, setting precision too small would result in futile attempts
247 * to syncronize to the unachievable precision.
249 * -6 is 1/64 sec, -7 is 1/128 sec and so on.
251 #define G_precision_exp -8
252 #define G_precision_sec (1.0 / (1 << (- G_precision_exp)))
254 /* Bool. After set to 1, never goes back to 0: */
255 smallint adjtimex_was_done;
256 smallint initial_poll_complete;
258 uint8_t discipline_state; // doc calls it c.state
259 uint8_t poll_exp; // s.poll
260 int polladj_count; // c.count
261 long kernel_freq_drift;
262 double last_update_offset; // c.last
263 double last_update_recv_time; // s.t
264 double discipline_jitter; // c.jitter
265 //TODO: add s.jitter - grep for it here and see clock_combine() in doc
266 #define USING_KERNEL_PLL_LOOP 1
267 #if !USING_KERNEL_PLL_LOOP
268 double discipline_freq_drift; // c.freq
269 //TODO: conditionally calculate wander? it's used only for logging
270 double discipline_wander; // c.wander
273 #define G (*ptr_to_globals)
275 static const int const_IPTOS_LOWDELAY = IPTOS_LOWDELAY;
278 #define VERB1 if (MAX_VERBOSE && G.verbose)
279 #define VERB2 if (MAX_VERBOSE >= 2 && G.verbose >= 2)
280 #define VERB3 if (MAX_VERBOSE >= 3 && G.verbose >= 3)
281 #define VERB4 if (MAX_VERBOSE >= 4 && G.verbose >= 4)
282 #define VERB5 if (MAX_VERBOSE >= 5 && G.verbose >= 5)
285 static double LOG2D(int a)
288 return 1.0 / (1UL << -a);
291 static ALWAYS_INLINE double SQUARE(double x)
295 static ALWAYS_INLINE double MAXD(double a, double b)
301 static ALWAYS_INLINE double MIND(double a, double b)
307 static NOINLINE double my_SQRT(double X)
314 double Xhalf = X * 0.5;
316 /* Fast and good approximation to 1/sqrt(X), black magic */
318 /*v.i = 0x5f3759df - (v.i >> 1);*/
319 v.i = 0x5f375a86 - (v.i >> 1); /* - this constant is slightly better */
320 invsqrt = v.f; /* better than 0.2% accuracy */
322 /* Refining it using Newton's method: x1 = x0 - f(x0)/f'(x0)
323 * f(x) = 1/(x*x) - X (f==0 when x = 1/sqrt(X))
325 * f(x)/f'(x) = (X - 1/(x*x)) / (2/(x*x*x)) = X*x*x*x/2 - x/2
326 * x1 = x0 - (X*x0*x0*x0/2 - x0/2) = 1.5*x0 - X*x0*x0*x0/2 = x0*(1.5 - (X/2)*x0*x0)
328 invsqrt = invsqrt * (1.5 - Xhalf * invsqrt * invsqrt); /* ~0.05% accuracy */
329 /* invsqrt = invsqrt * (1.5 - Xhalf * invsqrt * invsqrt); 2nd iter: ~0.0001% accuracy */
330 /* With 4 iterations, more than half results will be exact,
331 * at 6th iterations result stabilizes with about 72% results exact.
332 * We are well satisfied with 0.05% accuracy.
335 return X * invsqrt; /* X * 1/sqrt(X) ~= sqrt(X) */
337 static ALWAYS_INLINE double SQRT(double X)
339 /* If this arch doesn't use IEEE 754 floats, fall back to using libm */
340 if (sizeof(float) != 4)
343 /* This avoids needing libm, saves about 0.5k on x86-32 */
351 gettimeofday(&tv, NULL); /* never fails */
352 G.cur_time = tv.tv_sec + (1.0e-6 * tv.tv_usec) + OFFSET_1900_1970;
357 d_to_tv(double d, struct timeval *tv)
359 tv->tv_sec = (long)d;
360 tv->tv_usec = (d - tv->tv_sec) * 1000000;
364 lfp_to_d(l_fixedpt_t lfp)
367 lfp.int_partl = ntohl(lfp.int_partl);
368 lfp.fractionl = ntohl(lfp.fractionl);
369 ret = (double)lfp.int_partl + ((double)lfp.fractionl / UINT_MAX);
373 sfp_to_d(s_fixedpt_t sfp)
376 sfp.int_parts = ntohs(sfp.int_parts);
377 sfp.fractions = ntohs(sfp.fractions);
378 ret = (double)sfp.int_parts + ((double)sfp.fractions / USHRT_MAX);
381 #if ENABLE_FEATURE_NTPD_SERVER
386 lfp.int_partl = (uint32_t)d;
387 lfp.fractionl = (uint32_t)((d - lfp.int_partl) * UINT_MAX);
388 lfp.int_partl = htonl(lfp.int_partl);
389 lfp.fractionl = htonl(lfp.fractionl);
396 sfp.int_parts = (uint16_t)d;
397 sfp.fractions = (uint16_t)((d - sfp.int_parts) * USHRT_MAX);
398 sfp.int_parts = htons(sfp.int_parts);
399 sfp.fractions = htons(sfp.fractions);
405 dispersion(const datapoint_t *dp)
407 return dp->d_dispersion + FREQ_TOLERANCE * (G.cur_time - dp->d_recv_time);
411 root_distance(peer_t *p)
413 /* The root synchronization distance is the maximum error due to
414 * all causes of the local clock relative to the primary server.
415 * It is defined as half the total delay plus total dispersion
418 return MAXD(MINDISP, p->lastpkt_rootdelay + p->lastpkt_delay) / 2
419 + p->lastpkt_rootdisp
420 + p->filter_dispersion
421 + FREQ_TOLERANCE * (G.cur_time - p->lastpkt_recv_time)
426 set_next(peer_t *p, unsigned t)
428 p->next_action_time = G.cur_time + t;
432 * Peer clock filter and its helpers
435 filter_datapoints(peer_t *p)
439 double minoff, maxoff, wavg, sum, w;
440 double x = x; /* for compiler */
441 double oldest_off = oldest_off;
442 double oldest_age = oldest_age;
443 double newest_off = newest_off;
444 double newest_age = newest_age;
446 minoff = maxoff = p->filter_datapoint[0].d_offset;
447 for (i = 1; i < NUM_DATAPOINTS; i++) {
448 if (minoff > p->filter_datapoint[i].d_offset)
449 minoff = p->filter_datapoint[i].d_offset;
450 if (maxoff < p->filter_datapoint[i].d_offset)
451 maxoff = p->filter_datapoint[i].d_offset;
454 idx = p->datapoint_idx; /* most recent datapoint */
456 * Drop two outliers and take weighted average of the rest:
457 * most_recent/2 + older1/4 + older2/8 ... + older5/32 + older6/32
458 * we use older6/32, not older6/64 since sum of weights should be 1:
459 * 1/2 + 1/4 + 1/8 + 1/16 + 1/32 + 1/32 = 1
465 * filter_dispersion = \ -------------
472 for (i = 0; i < NUM_DATAPOINTS; i++) {
474 bb_error_msg("datapoint[%d]: off:%f disp:%f(%f) age:%f%s",
476 p->filter_datapoint[idx].d_offset,
477 p->filter_datapoint[idx].d_dispersion, dispersion(&p->filter_datapoint[idx]),
478 G.cur_time - p->filter_datapoint[idx].d_recv_time,
479 (minoff == p->filter_datapoint[idx].d_offset || maxoff == p->filter_datapoint[idx].d_offset)
480 ? " (outlier by offset)" : ""
484 sum += dispersion(&p->filter_datapoint[idx]) / (2 << i);
486 if (minoff == p->filter_datapoint[idx].d_offset) {
487 minoff -= 1; /* so that we don't match it ever again */
489 if (maxoff == p->filter_datapoint[idx].d_offset) {
492 oldest_off = p->filter_datapoint[idx].d_offset;
493 oldest_age = G.cur_time - p->filter_datapoint[idx].d_recv_time;
496 newest_off = oldest_off;
497 newest_age = oldest_age;
504 idx = (idx - 1) & (NUM_DATAPOINTS - 1);
506 p->filter_dispersion = sum;
507 wavg += x; /* add another older6/64 to form older6/32 */
508 /* Fix systematic underestimation with large poll intervals.
509 * Imagine that we still have a bit of uncorrected drift,
510 * and poll interval is big (say, 100 sec). Offsets form a progression:
511 * 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 - 0.7 is most recent.
512 * The algorithm above drops 0.0 and 0.7 as outliers,
513 * and then we have this estimation, ~25% off from 0.7:
514 * 0.1/32 + 0.2/32 + 0.3/16 + 0.4/8 + 0.5/4 + 0.6/2 = 0.503125
516 x = oldest_age - newest_age;
518 x = newest_age / x; /* in above example, 100 / (600 - 100) */
519 if (x < 1) { /* paranoia check */
520 x = (newest_off - oldest_off) * x; /* 0.5 * 100/500 = 0.1 */
524 p->filter_offset = wavg;
526 /* +----- -----+ ^ 1/2
530 * filter_jitter = | --- * / (avg-offset_j) |
534 * where n is the number of valid datapoints in the filter (n > 1);
535 * if filter_jitter < precision then filter_jitter = precision
538 for (i = 0; i < NUM_DATAPOINTS; i++) {
539 sum += SQUARE(wavg - p->filter_datapoint[i].d_offset);
541 sum = SQRT(sum / NUM_DATAPOINTS);
542 p->filter_jitter = sum > G_precision_sec ? sum : G_precision_sec;
544 VERB3 bb_error_msg("filter offset:%f(corr:%e) disp:%f jitter:%f",
546 p->filter_dispersion,
552 reset_peer_stats(peer_t *p, double offset)
555 for (i = 0; i < NUM_DATAPOINTS; i++) {
556 if (offset < 16 * STEP_THRESHOLD) {
557 p->filter_datapoint[i].d_recv_time -= offset;
558 if (p->filter_datapoint[i].d_offset != 0) {
559 p->filter_datapoint[i].d_offset -= offset;
562 p->filter_datapoint[i].d_recv_time = G.cur_time;
563 p->filter_datapoint[i].d_offset = 0;
564 p->filter_datapoint[i].d_dispersion = MAXDISP;
567 if (offset < 16 * STEP_THRESHOLD) {
568 p->lastpkt_recv_time -= offset;
570 p->reachable_bits = 0;
571 p->lastpkt_recv_time = G.cur_time;
573 filter_datapoints(p); /* recalc p->filter_xxx */
574 p->next_action_time -= offset;
575 VERB5 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
583 p = xzalloc(sizeof(*p));
584 p->p_lsa = xhost2sockaddr(s, 123);
585 p->p_dotted = xmalloc_sockaddr2dotted_noport(&p->p_lsa->u.sa);
587 p->p_xmt_msg.m_status = MODE_CLIENT | (NTP_VERSION << 3);
588 p->next_action_time = G.cur_time; /* = set_next(p, 0); */
589 reset_peer_stats(p, 16 * STEP_THRESHOLD);
591 llist_add_to(&G.ntp_peers, p);
597 const struct sockaddr *from, const struct sockaddr *to, socklen_t addrlen,
598 msg_t *msg, ssize_t len)
604 ret = sendto(fd, msg, len, MSG_DONTWAIT, to, addrlen);
606 ret = send_to_from(fd, msg, len, MSG_DONTWAIT, to, from, addrlen);
609 bb_perror_msg("send failed");
616 send_query_to_peer(peer_t *p)
618 /* Why do we need to bind()?
619 * See what happens when we don't bind:
621 * socket(PF_INET, SOCK_DGRAM, IPPROTO_IP) = 3
622 * setsockopt(3, SOL_IP, IP_TOS, [16], 4) = 0
623 * gettimeofday({1259071266, 327885}, NULL) = 0
624 * sendto(3, "xxx", 48, MSG_DONTWAIT, {sa_family=AF_INET, sin_port=htons(123), sin_addr=inet_addr("10.34.32.125")}, 16) = 48
625 * ^^^ we sent it from some source port picked by kernel.
626 * time(NULL) = 1259071266
627 * write(2, "ntpd: entering poll 15 secs\n", 28) = 28
628 * poll([{fd=3, events=POLLIN}], 1, 15000) = 1 ([{fd=3, revents=POLLIN}])
629 * recv(3, "yyy", 68, MSG_DONTWAIT) = 48
630 * ^^^ this recv will receive packets to any local port!
632 * Uncomment this and use strace to see it in action:
634 #define PROBE_LOCAL_ADDR /* { len_and_sockaddr lsa; lsa.len = LSA_SIZEOF_SA; getsockname(p->query.fd, &lsa.u.sa, &lsa.len); } */
638 len_and_sockaddr *local_lsa;
640 family = p->p_lsa->u.sa.sa_family;
641 p->p_fd = fd = xsocket_type(&local_lsa, family, SOCK_DGRAM);
642 /* local_lsa has "null" address and port 0 now.
643 * bind() ensures we have a *particular port* selected by kernel
644 * and remembered in p->p_fd, thus later recv(p->p_fd)
645 * receives only packets sent to this port.
648 xbind(fd, &local_lsa->u.sa, local_lsa->len);
650 #if ENABLE_FEATURE_IPV6
651 if (family == AF_INET)
653 setsockopt(fd, IPPROTO_IP, IP_TOS, &const_IPTOS_LOWDELAY, sizeof(const_IPTOS_LOWDELAY));
658 * Send out a random 64-bit number as our transmit time. The NTP
659 * server will copy said number into the originate field on the
660 * response that it sends us. This is totally legal per the SNTP spec.
662 * The impact of this is two fold: we no longer send out the current
663 * system time for the world to see (which may aid an attacker), and
664 * it gives us a (not very secure) way of knowing that we're not
665 * getting spoofed by an attacker that can't capture our traffic
666 * but can spoof packets from the NTP server we're communicating with.
668 * Save the real transmit timestamp locally.
670 p->p_xmt_msg.m_xmttime.int_partl = random();
671 p->p_xmt_msg.m_xmttime.fractionl = random();
672 p->p_xmttime = gettime1900d();
674 if (do_sendto(p->p_fd, /*from:*/ NULL, /*to:*/ &p->p_lsa->u.sa, /*addrlen:*/ p->p_lsa->len,
675 &p->p_xmt_msg, NTP_MSGSIZE_NOAUTH) == -1
679 set_next(p, RETRY_INTERVAL);
683 p->reachable_bits <<= 1;
684 VERB1 bb_error_msg("sent query to %s", p->p_dotted);
685 set_next(p, RESPONSE_INTERVAL);
689 static void run_script(const char *action)
697 argv[0] = (char*) G.script_name;
698 argv[1] = (char*) action;
701 VERB1 bb_error_msg("executing '%s %s'", G.script_name, action);
703 env1 = xasprintf("stratum=%u", G.stratum);
705 env2 = xasprintf("freq_drift_ppm=%ld", G.kernel_freq_drift);
707 /* Other items of potential interest: selected peer,
708 * rootdelay, reftime, rootdisp, refid, ntp_status, poll_exp,
709 * last_update_offset, last_update_recv_time, discipline_jitter
712 /* Don't want to wait: it may run hwclock --systohc, and that
713 * may take some time (seconds): */
714 /*wait4pid(spawn(argv));*/
716 G.last_script_run = G.cur_time;
719 unsetenv("freq_drift_ppm");
725 step_time(double offset)
733 gettimeofday(&tv, NULL); /* never fails */
734 dtime = offset + tv.tv_sec;
735 dtime += 1.0e-6 * tv.tv_usec;
738 if (settimeofday(&tv, NULL) == -1)
739 bb_perror_msg_and_die("settimeofday");
742 strftime(buf, sizeof(buf), "%a %b %e %H:%M:%S %Z %Y", localtime(&tval));
744 bb_error_msg("setting clock to %s (offset %fs)", buf, offset);
746 /* Correct various fields which contain time-relative values: */
748 /* p->lastpkt_recv_time, p->next_action_time and such: */
749 for (item = G.ntp_peers; item != NULL; item = item->link) {
750 peer_t *pp = (peer_t *) item->data;
751 reset_peer_stats(pp, offset);
754 G.cur_time -= offset;
755 G.last_update_recv_time -= offset;
760 * Selection and clustering, and their helpers
768 compare_point_edge(const void *aa, const void *bb)
770 const point_t *a = aa;
771 const point_t *b = bb;
772 if (a->edge < b->edge) {
775 return (a->edge > b->edge);
782 compare_survivor_metric(const void *aa, const void *bb)
784 const survivor_t *a = aa;
785 const survivor_t *b = bb;
786 if (a->metric < b->metric) {
789 return (a->metric > b->metric);
792 fit(peer_t *p, double rd)
794 if ((p->reachable_bits & (p->reachable_bits-1)) == 0) {
795 /* One or zero bits in reachable_bits */
796 VERB3 bb_error_msg("peer %s unfit for selection: unreachable", p->p_dotted);
799 #if 0 /* we filter out such packets earlier */
800 if ((p->lastpkt_status & LI_ALARM) == LI_ALARM
801 || p->lastpkt_stratum >= MAXSTRAT
803 VERB3 bb_error_msg("peer %s unfit for selection: bad status/stratum", p->p_dotted);
807 /* rd is root_distance(p) */
808 if (rd > MAXDIST + FREQ_TOLERANCE * (1 << G.poll_exp)) {
809 VERB3 bb_error_msg("peer %s unfit for selection: root distance too high", p->p_dotted);
813 // /* Do we have a loop? */
814 // if (p->refid == p->dstaddr || p->refid == s.refid)
819 select_and_cluster(void)
823 int size = 3 * G.peer_cnt;
824 /* for selection algorithm */
826 unsigned num_points, num_candidates;
828 unsigned num_falsetickers;
829 /* for cluster algorithm */
830 survivor_t survivor[size];
831 unsigned num_survivors;
837 if (G.initial_poll_complete) while (item != NULL) {
838 peer_t *p = (peer_t *) item->data;
839 double rd = root_distance(p);
840 double offset = p->filter_offset;
847 VERB4 bb_error_msg("interval: [%f %f %f] %s",
853 point[num_points].p = p;
854 point[num_points].type = -1;
855 point[num_points].edge = offset - rd;
857 point[num_points].p = p;
858 point[num_points].type = 0;
859 point[num_points].edge = offset;
861 point[num_points].p = p;
862 point[num_points].type = 1;
863 point[num_points].edge = offset + rd;
867 num_candidates = num_points / 3;
868 if (num_candidates == 0) {
869 VERB3 bb_error_msg("no valid datapoints, no peer selected");
872 //TODO: sorting does not seem to be done in reference code
873 qsort(point, num_points, sizeof(point[0]), compare_point_edge);
875 /* Start with the assumption that there are no falsetickers.
876 * Attempt to find a nonempty intersection interval containing
877 * the midpoints of all truechimers.
878 * If a nonempty interval cannot be found, increase the number
879 * of assumed falsetickers by one and try again.
880 * If a nonempty interval is found and the number of falsetickers
881 * is less than the number of truechimers, a majority has been found
882 * and the midpoint of each truechimer represents
883 * the candidates available to the cluster algorithm.
885 num_falsetickers = 0;
888 unsigned num_midpoints = 0;
893 for (i = 0; i < num_points; i++) {
895 * if (point[i].type == -1) c++;
896 * if (point[i].type == 1) c--;
897 * and it's simpler to do it this way:
900 if (c >= num_candidates - num_falsetickers) {
901 /* If it was c++ and it got big enough... */
905 if (point[i].type == 0)
909 for (i = num_points-1; i >= 0; i--) {
911 if (c >= num_candidates - num_falsetickers) {
912 high = point[i].edge;
915 if (point[i].type == 0)
918 /* If the number of midpoints is greater than the number
919 * of allowed falsetickers, the intersection contains at
920 * least one truechimer with no midpoint - bad.
921 * Also, interval should be nonempty.
923 if (num_midpoints <= num_falsetickers && low < high)
926 if (num_falsetickers * 2 >= num_candidates) {
927 VERB3 bb_error_msg("too many falsetickers:%d (candidates:%d), no peer selected",
928 num_falsetickers, num_candidates);
932 VERB3 bb_error_msg("selected interval: [%f, %f]; candidates:%d falsetickers:%d",
933 low, high, num_candidates, num_falsetickers);
937 /* Construct a list of survivors (p, metric)
938 * from the chime list, where metric is dominated
939 * first by stratum and then by root distance.
940 * All other things being equal, this is the order of preference.
943 for (i = 0; i < num_points; i++) {
946 if (point[i].edge < low || point[i].edge > high)
949 survivor[num_survivors].p = p;
950 //TODO: save root_distance in point_t and reuse here?
951 survivor[num_survivors].metric = MAXDIST * p->lastpkt_stratum + root_distance(p);
952 VERB4 bb_error_msg("survivor[%d] metric:%f peer:%s",
953 num_survivors, survivor[num_survivors].metric, p->p_dotted);
956 /* There must be at least MIN_SELECTED survivors to satisfy the
957 * correctness assertions. Ordinarily, the Byzantine criteria
958 * require four survivors, but for the demonstration here, one
961 if (num_survivors < MIN_SELECTED) {
962 VERB3 bb_error_msg("num_survivors %d < %d, no peer selected",
963 num_survivors, MIN_SELECTED);
967 //looks like this is ONLY used by the fact that later we pick survivor[0].
968 //we can avoid sorting then, just find the minimum once!
969 qsort(survivor, num_survivors, sizeof(survivor[0]), compare_survivor_metric);
971 /* For each association p in turn, calculate the selection
972 * jitter p->sjitter as the square root of the sum of squares
973 * (p->offset - q->offset) over all q associations. The idea is
974 * to repeatedly discard the survivor with maximum selection
975 * jitter until a termination condition is met.
978 unsigned max_idx = max_idx;
979 double max_selection_jitter = max_selection_jitter;
980 double min_jitter = min_jitter;
982 if (num_survivors <= MIN_CLUSTERED) {
983 VERB3 bb_error_msg("num_survivors %d <= %d, not discarding more",
984 num_survivors, MIN_CLUSTERED);
988 /* To make sure a few survivors are left
989 * for the clustering algorithm to chew on,
990 * we stop if the number of survivors
991 * is less than or equal to MIN_CLUSTERED (3).
993 for (i = 0; i < num_survivors; i++) {
994 double selection_jitter_sq;
995 peer_t *p = survivor[i].p;
997 if (i == 0 || p->filter_jitter < min_jitter)
998 min_jitter = p->filter_jitter;
1000 selection_jitter_sq = 0;
1001 for (j = 0; j < num_survivors; j++) {
1002 peer_t *q = survivor[j].p;
1003 selection_jitter_sq += SQUARE(p->filter_offset - q->filter_offset);
1005 if (i == 0 || selection_jitter_sq > max_selection_jitter) {
1006 max_selection_jitter = selection_jitter_sq;
1009 VERB5 bb_error_msg("survivor %d selection_jitter^2:%f",
1010 i, selection_jitter_sq);
1012 max_selection_jitter = SQRT(max_selection_jitter / num_survivors);
1013 VERB4 bb_error_msg("max_selection_jitter (at %d):%f min_jitter:%f",
1014 max_idx, max_selection_jitter, min_jitter);
1016 /* If the maximum selection jitter is less than the
1017 * minimum peer jitter, then tossing out more survivors
1018 * will not lower the minimum peer jitter, so we might
1021 if (max_selection_jitter < min_jitter) {
1022 VERB3 bb_error_msg("max_selection_jitter:%f < min_jitter:%f, num_survivors:%d, not discarding more",
1023 max_selection_jitter, min_jitter, num_survivors);
1027 /* Delete survivor[max_idx] from the list
1028 * and go around again.
1030 VERB5 bb_error_msg("dropping survivor %d", max_idx);
1032 while (max_idx < num_survivors) {
1033 survivor[max_idx] = survivor[max_idx + 1];
1038 /* Pick the best clock. If the old system peer is on the list
1039 * and at the same stratum as the first survivor on the list,
1040 * then don't do a clock hop. Otherwise, select the first
1041 * survivor on the list as the new system peer.
1043 //TODO - see clock_combine()
1044 VERB3 bb_error_msg("selected peer %s filter_offset:%f age:%f",
1045 survivor[0].p->p_dotted,
1046 survivor[0].p->filter_offset,
1047 G.cur_time - survivor[0].p->lastpkt_recv_time
1049 return survivor[0].p;
1054 * Local clock discipline and its helpers
1057 set_new_values(int disc_state, double offset, double recv_time)
1059 /* Enter new state and set state variables. Note we use the time
1060 * of the last clock filter sample, which must be earlier than
1063 VERB3 bb_error_msg("disc_state=%d last update offset=%f recv_time=%f",
1064 disc_state, offset, recv_time);
1065 G.discipline_state = disc_state;
1066 G.last_update_offset = offset;
1067 G.last_update_recv_time = recv_time;
1069 /* Clock state definitions */
1070 #define STATE_NSET 0 /* initial state, "nothing is set" */
1071 #define STATE_FSET 1 /* frequency set from file */
1072 #define STATE_SPIK 2 /* spike detected */
1073 #define STATE_FREQ 3 /* initial frequency */
1074 #define STATE_SYNC 4 /* clock synchronized (normal operation) */
1075 /* Return: -1: decrease poll interval, 0: leave as is, 1: increase */
1077 update_local_clock(peer_t *p)
1080 long old_tmx_offset;
1082 double offset = p->filter_offset;
1083 double recv_time = p->lastpkt_recv_time;
1085 #if !USING_KERNEL_PLL_LOOP
1088 double since_last_update;
1089 double etemp, dtemp;
1091 abs_offset = fabs(offset);
1093 /* If the offset is too large, give up and go home */
1094 if (abs_offset > PANIC_THRESHOLD) {
1095 bb_error_msg_and_die("offset %f far too big, exiting", offset);
1098 /* If this is an old update, for instance as the result
1099 * of a system peer change, avoid it. We never use
1100 * an old sample or the same sample twice.
1102 if (recv_time <= G.last_update_recv_time) {
1103 VERB3 bb_error_msg("same or older datapoint: %f >= %f, not using it",
1104 G.last_update_recv_time, recv_time);
1105 return 0; /* "leave poll interval as is" */
1108 /* Clock state machine transition function. This is where the
1109 * action is and defines how the system reacts to large time
1110 * and frequency errors.
1112 since_last_update = recv_time - G.reftime;
1113 #if !USING_KERNEL_PLL_LOOP
1116 if (G.discipline_state == STATE_FREQ) {
1117 /* Ignore updates until the stepout threshold */
1118 if (since_last_update < WATCH_THRESHOLD) {
1119 VERB3 bb_error_msg("measuring drift, datapoint ignored, %f sec remains",
1120 WATCH_THRESHOLD - since_last_update);
1121 return 0; /* "leave poll interval as is" */
1123 #if !USING_KERNEL_PLL_LOOP
1124 freq_drift = (offset - G.last_update_offset) / since_last_update;
1128 /* There are two main regimes: when the
1129 * offset exceeds the step threshold and when it does not.
1131 if (abs_offset > STEP_THRESHOLD) {
1132 switch (G.discipline_state) {
1134 /* The first outlyer: ignore it, switch to SPIK state */
1135 VERB3 bb_error_msg("offset:%f - spike detected", offset);
1136 G.discipline_state = STATE_SPIK;
1137 return -1; /* "decrease poll interval" */
1140 /* Ignore succeeding outlyers until either an inlyer
1141 * is found or the stepout threshold is exceeded.
1143 if (since_last_update < WATCH_THRESHOLD) {
1144 VERB3 bb_error_msg("spike detected, datapoint ignored, %f sec remains",
1145 WATCH_THRESHOLD - since_last_update);
1146 return -1; /* "decrease poll interval" */
1148 /* fall through: we need to step */
1151 /* Step the time and clamp down the poll interval.
1153 * In NSET state an initial frequency correction is
1154 * not available, usually because the frequency file has
1155 * not yet been written. Since the time is outside the
1156 * capture range, the clock is stepped. The frequency
1157 * will be set directly following the stepout interval.
1159 * In FSET state the initial frequency has been set
1160 * from the frequency file. Since the time is outside
1161 * the capture range, the clock is stepped immediately,
1162 * rather than after the stepout interval. Guys get
1163 * nervous if it takes 17 minutes to set the clock for
1166 * In SPIK state the stepout threshold has expired and
1167 * the phase is still above the step threshold. Note
1168 * that a single spike greater than the step threshold
1169 * is always suppressed, even at the longer poll
1172 VERB3 bb_error_msg("stepping time by %f; poll_exp=MINPOLL", offset);
1174 if (option_mask32 & OPT_q) {
1175 /* We were only asked to set time once. Done. */
1179 G.polladj_count = 0;
1180 G.poll_exp = MINPOLL;
1181 G.stratum = MAXSTRAT;
1185 if (G.discipline_state == STATE_NSET) {
1186 set_new_values(STATE_FREQ, /*offset:*/ 0, recv_time);
1187 return 1; /* "ok to increase poll interval" */
1189 set_new_values(STATE_SYNC, /*offset:*/ 0, recv_time);
1191 } else { /* abs_offset <= STEP_THRESHOLD */
1193 if (G.poll_exp < MINPOLL && G.initial_poll_complete) {
1194 VERB3 bb_error_msg("small offset:%f, disabling burst mode", offset);
1195 G.polladj_count = 0;
1196 G.poll_exp = MINPOLL;
1199 /* Compute the clock jitter as the RMS of exponentially
1200 * weighted offset differences. Used by the poll adjust code.
1202 etemp = SQUARE(G.discipline_jitter);
1203 dtemp = SQUARE(MAXD(fabs(offset - G.last_update_offset), G_precision_sec));
1204 G.discipline_jitter = SQRT(etemp + (dtemp - etemp) / AVG);
1205 VERB3 bb_error_msg("discipline jitter=%f", G.discipline_jitter);
1207 switch (G.discipline_state) {
1209 if (option_mask32 & OPT_q) {
1210 /* We were only asked to set time once.
1211 * The clock is precise enough, no need to step.
1215 /* This is the first update received and the frequency
1216 * has not been initialized. The first thing to do
1217 * is directly measure the oscillator frequency.
1219 set_new_values(STATE_FREQ, offset, recv_time);
1220 VERB3 bb_error_msg("transitioning to FREQ, datapoint ignored");
1221 return 0; /* "leave poll interval as is" */
1223 #if 0 /* this is dead code for now */
1225 /* This is the first update and the frequency
1226 * has been initialized. Adjust the phase, but
1227 * don't adjust the frequency until the next update.
1229 set_new_values(STATE_SYNC, offset, recv_time);
1230 /* freq_drift remains 0 */
1235 /* since_last_update >= WATCH_THRESHOLD, we waited enough.
1236 * Correct the phase and frequency and switch to SYNC state.
1237 * freq_drift was already estimated (see code above)
1239 set_new_values(STATE_SYNC, offset, recv_time);
1243 #if !USING_KERNEL_PLL_LOOP
1244 /* Compute freq_drift due to PLL and FLL contributions.
1246 * The FLL and PLL frequency gain constants
1247 * depend on the poll interval and Allan
1248 * intercept. The FLL is not used below one-half
1249 * the Allan intercept. Above that the loop gain
1250 * increases in steps to 1 / AVG.
1252 if ((1 << G.poll_exp) > ALLAN / 2) {
1253 etemp = FLL - G.poll_exp;
1256 freq_drift += (offset - G.last_update_offset) / (MAXD(since_last_update, ALLAN) * etemp);
1258 /* For the PLL the integration interval
1259 * (numerator) is the minimum of the update
1260 * interval and poll interval. This allows
1261 * oversampling, but not undersampling.
1263 etemp = MIND(since_last_update, (1 << G.poll_exp));
1264 dtemp = (4 * PLL) << G.poll_exp;
1265 freq_drift += offset * etemp / SQUARE(dtemp);
1267 set_new_values(STATE_SYNC, offset, recv_time);
1270 if (G.stratum != p->lastpkt_stratum + 1) {
1271 G.stratum = p->lastpkt_stratum + 1;
1272 run_script("stratum");
1276 G.reftime = G.cur_time;
1277 G.ntp_status = p->lastpkt_status;
1278 G.refid = p->lastpkt_refid;
1279 G.rootdelay = p->lastpkt_rootdelay + p->lastpkt_delay;
1280 dtemp = p->filter_jitter; // SQRT(SQUARE(p->filter_jitter) + SQUARE(s.jitter));
1281 dtemp += MAXD(p->filter_dispersion + FREQ_TOLERANCE * (G.cur_time - p->lastpkt_recv_time) + abs_offset, MINDISP);
1282 G.rootdisp = p->lastpkt_rootdisp + dtemp;
1283 VERB3 bb_error_msg("updating leap/refid/reftime/rootdisp from peer %s", p->p_dotted);
1285 /* We are in STATE_SYNC now, but did not do adjtimex yet.
1286 * (Any other state does not reach this, they all return earlier)
1287 * By this time, freq_drift and G.last_update_offset are set
1288 * to values suitable for adjtimex.
1290 #if !USING_KERNEL_PLL_LOOP
1291 /* Calculate the new frequency drift and frequency stability (wander).
1292 * Compute the clock wander as the RMS of exponentially weighted
1293 * frequency differences. This is not used directly, but can,
1294 * along with the jitter, be a highly useful monitoring and
1297 dtemp = G.discipline_freq_drift + freq_drift;
1298 G.discipline_freq_drift = MAXD(MIND(MAXDRIFT, dtemp), -MAXDRIFT);
1299 etemp = SQUARE(G.discipline_wander);
1300 dtemp = SQUARE(dtemp);
1301 G.discipline_wander = SQRT(etemp + (dtemp - etemp) / AVG);
1303 VERB3 bb_error_msg("discipline freq_drift=%.9f(int:%ld corr:%e) wander=%f",
1304 G.discipline_freq_drift,
1305 (long)(G.discipline_freq_drift * 65536e6),
1307 G.discipline_wander);
1310 memset(&tmx, 0, sizeof(tmx));
1311 if (adjtimex(&tmx) < 0)
1312 bb_perror_msg_and_die("adjtimex");
1313 VERB3 bb_error_msg("p adjtimex freq:%ld offset:%ld constant:%ld status:0x%x",
1314 tmx.freq, tmx.offset, tmx.constant, tmx.status);
1318 if (!G.adjtimex_was_done) {
1319 G.adjtimex_was_done = 1;
1320 /* When we use adjtimex for the very first time,
1321 * we need to ADD to pre-existing tmx.offset - it may be !0
1323 memset(&tmx, 0, sizeof(tmx));
1324 if (adjtimex(&tmx) < 0)
1325 bb_perror_msg_and_die("adjtimex");
1326 old_tmx_offset = tmx.offset;
1328 memset(&tmx, 0, sizeof(tmx));
1330 //doesn't work, offset remains 0 (!) in kernel:
1331 //ntpd: set adjtimex freq:1786097 tmx.offset:77487
1332 //ntpd: prev adjtimex freq:1786097 tmx.offset:0
1333 //ntpd: cur adjtimex freq:1786097 tmx.offset:0
1334 tmx.modes = ADJ_FREQUENCY | ADJ_OFFSET;
1335 /* 65536 is one ppm */
1336 tmx.freq = G.discipline_freq_drift * 65536e6;
1337 tmx.offset = G.last_update_offset * 1000000; /* usec */
1339 tmx.modes = ADJ_OFFSET | ADJ_STATUS | ADJ_TIMECONST;// | ADJ_MAXERROR | ADJ_ESTERROR;
1340 tmx.offset = (G.last_update_offset * 1000000) /* usec */
1341 /* + (G.last_update_offset < 0 ? -0.5 : 0.5) - too small to bother */
1342 + old_tmx_offset; /* almost always 0 */
1343 tmx.status = STA_PLL;
1344 if (G.ntp_status & LI_PLUSSEC)
1345 tmx.status |= STA_INS;
1346 if (G.ntp_status & LI_MINUSSEC)
1347 tmx.status |= STA_DEL;
1348 tmx.constant = G.poll_exp - 4;
1349 //tmx.esterror = (u_int32)(clock_jitter * 1e6);
1350 //tmx.maxerror = (u_int32)((sys_rootdelay / 2 + sys_rootdisp) * 1e6);
1351 rc = adjtimex(&tmx);
1353 bb_perror_msg_and_die("adjtimex");
1354 /* NB: here kernel returns constant == G.poll_exp, not == G.poll_exp - 4.
1355 * Not sure why. Perhaps it is normal.
1357 VERB3 bb_error_msg("adjtimex:%d freq:%ld offset:%ld constant:%ld status:0x%x",
1358 rc, tmx.freq, tmx.offset, tmx.constant, tmx.status);
1361 /* always gives the same output as above msg */
1362 memset(&tmx, 0, sizeof(tmx));
1363 if (adjtimex(&tmx) < 0)
1364 bb_perror_msg_and_die("adjtimex");
1365 VERB3 bb_error_msg("c adjtimex freq:%ld offset:%ld constant:%ld status:0x%x",
1366 tmx.freq, tmx.offset, tmx.constant, tmx.status);
1369 if (G.kernel_freq_drift != tmx.freq / 65536) {
1370 G.kernel_freq_drift = tmx.freq / 65536;
1371 VERB2 bb_error_msg("kernel clock drift: %ld ppm", G.kernel_freq_drift);
1374 return 1; /* "ok to increase poll interval" */
1379 * We've got a new reply packet from a peer, process it
1383 retry_interval(void)
1385 /* Local problem, want to retry soon */
1386 unsigned interval, r;
1387 interval = RETRY_INTERVAL;
1389 interval += r % (unsigned)(RETRY_INTERVAL / 4);
1390 VERB3 bb_error_msg("chose retry interval:%u", interval);
1394 poll_interval(int exponent)
1396 unsigned interval, r;
1397 exponent = G.poll_exp + exponent;
1400 interval = 1 << exponent;
1402 interval += ((r & (interval-1)) >> 4) + ((r >> 8) & 1); /* + 1/16 of interval, max */
1403 VERB3 bb_error_msg("chose poll interval:%u (poll_exp:%d exp:%d)", interval, G.poll_exp, exponent);
1406 static NOINLINE void
1407 recv_and_process_peer_pkt(peer_t *p)
1412 double T1, T2, T3, T4;
1414 datapoint_t *datapoint;
1417 /* We can recvfrom here and check from.IP, but some multihomed
1418 * ntp servers reply from their *other IP*.
1419 * TODO: maybe we should check at least what we can: from.port == 123?
1421 size = recv(p->p_fd, &msg, sizeof(msg), MSG_DONTWAIT);
1423 bb_perror_msg("recv(%s) error", p->p_dotted);
1424 if (errno == EHOSTUNREACH || errno == EHOSTDOWN
1425 || errno == ENETUNREACH || errno == ENETDOWN
1426 || errno == ECONNREFUSED || errno == EADDRNOTAVAIL
1429 //TODO: always do this?
1430 interval = retry_interval();
1431 goto set_next_and_close_sock;
1436 if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
1437 bb_error_msg("malformed packet received from %s", p->p_dotted);
1441 if (msg.m_orgtime.int_partl != p->p_xmt_msg.m_xmttime.int_partl
1442 || msg.m_orgtime.fractionl != p->p_xmt_msg.m_xmttime.fractionl
1447 if ((msg.m_status & LI_ALARM) == LI_ALARM
1448 || msg.m_stratum == 0
1449 || msg.m_stratum > NTP_MAXSTRATUM
1451 // TODO: stratum 0 responses may have commands in 32-bit m_refid field:
1452 // "DENY", "RSTR" - peer does not like us at all
1453 // "RATE" - peer is overloaded, reduce polling freq
1454 interval = poll_interval(0);
1455 bb_error_msg("reply from %s: not synced, next query in %us", p->p_dotted, interval);
1456 goto set_next_and_close_sock;
1459 // /* Verify valid root distance */
1460 // if (msg.m_rootdelay / 2 + msg.m_rootdisp >= MAXDISP || p->lastpkt_reftime > msg.m_xmt)
1461 // return; /* invalid header values */
1463 p->lastpkt_status = msg.m_status;
1464 p->lastpkt_stratum = msg.m_stratum;
1465 p->lastpkt_rootdelay = sfp_to_d(msg.m_rootdelay);
1466 p->lastpkt_rootdisp = sfp_to_d(msg.m_rootdisp);
1467 p->lastpkt_refid = msg.m_refid;
1470 * From RFC 2030 (with a correction to the delay math):
1472 * Timestamp Name ID When Generated
1473 * ------------------------------------------------------------
1474 * Originate Timestamp T1 time request sent by client
1475 * Receive Timestamp T2 time request received by server
1476 * Transmit Timestamp T3 time reply sent by server
1477 * Destination Timestamp T4 time reply received by client
1479 * The roundtrip delay and local clock offset are defined as
1481 * delay = (T4 - T1) - (T3 - T2); offset = ((T2 - T1) + (T3 - T4)) / 2
1484 T2 = lfp_to_d(msg.m_rectime);
1485 T3 = lfp_to_d(msg.m_xmttime);
1488 p->lastpkt_recv_time = T4;
1490 VERB5 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
1491 p->datapoint_idx = p->reachable_bits ? (p->datapoint_idx + 1) % NUM_DATAPOINTS : 0;
1492 datapoint = &p->filter_datapoint[p->datapoint_idx];
1493 datapoint->d_recv_time = T4;
1494 datapoint->d_offset = ((T2 - T1) + (T3 - T4)) / 2;
1495 /* The delay calculation is a special case. In cases where the
1496 * server and client clocks are running at different rates and
1497 * with very fast networks, the delay can appear negative. In
1498 * order to avoid violating the Principle of Least Astonishment,
1499 * the delay is clamped not less than the system precision.
1501 p->lastpkt_delay = (T4 - T1) - (T3 - T2);
1502 if (p->lastpkt_delay < G_precision_sec)
1503 p->lastpkt_delay = G_precision_sec;
1504 datapoint->d_dispersion = LOG2D(msg.m_precision_exp) + G_precision_sec;
1505 if (!p->reachable_bits) {
1506 /* 1st datapoint ever - replicate offset in every element */
1508 for (i = 1; i < NUM_DATAPOINTS; i++) {
1509 p->filter_datapoint[i].d_offset = datapoint->d_offset;
1513 p->reachable_bits |= 1;
1514 if ((MAX_VERBOSE && G.verbose) || (option_mask32 & OPT_w)) {
1515 bb_error_msg("reply from %s: reach 0x%02x offset %f delay %f status 0x%02x strat %d refid 0x%08x rootdelay %f",
1518 datapoint->d_offset,
1523 p->lastpkt_rootdelay
1524 /* not shown: m_ppoll, m_precision_exp, m_rootdisp,
1525 * m_reftime, m_orgtime, m_rectime, m_xmttime
1530 /* Muck with statictics and update the clock */
1531 filter_datapoints(p);
1532 q = select_and_cluster();
1536 if (!(option_mask32 & OPT_w))
1537 rc = update_local_clock(q);
1541 /* Adjust the poll interval by comparing the current offset
1542 * with the clock jitter. If the offset is less than
1543 * the clock jitter times a constant, then the averaging interval
1544 * is increased, otherwise it is decreased. A bit of hysteresis
1545 * helps calm the dance. Works best using burst mode.
1548 bb_error_msg("offset:%f POLLADJ_GATE*discipline_jitter:%f poll:%s",
1549 q->filter_offset, POLLADJ_GATE * G.discipline_jitter,
1550 fabs(q->filter_offset) < POLLADJ_GATE * G.discipline_jitter
1554 if (rc > 0 && fabs(q->filter_offset) < POLLADJ_GATE * G.discipline_jitter) {
1555 /* was += G.poll_exp but it is a bit
1556 * too optimistic for my taste at high poll_exp's */
1557 G.polladj_count += MINPOLL;
1558 if (G.polladj_count > POLLADJ_LIMIT) {
1559 G.polladj_count = 0;
1560 if (G.poll_exp < MAXPOLL) {
1562 VERB3 bb_error_msg("polladj: discipline_jitter:%f ++poll_exp=%d",
1563 G.discipline_jitter, G.poll_exp);
1566 VERB3 bb_error_msg("polladj: incr:%d", G.polladj_count);
1569 G.polladj_count -= G.poll_exp * 2;
1570 if (G.polladj_count < -POLLADJ_LIMIT) {
1571 G.polladj_count = 0;
1572 if (G.poll_exp > MINPOLL) {
1576 /* Correct p->next_action_time in each peer
1577 * which waits for sending, so that they send earlier.
1578 * Old pp->next_action_time are on the order
1579 * of t + (1 << old_poll_exp) + small_random,
1580 * we simply need to subtract ~half of that.
1582 for (item = G.ntp_peers; item != NULL; item = item->link) {
1583 peer_t *pp = (peer_t *) item->data;
1585 pp->next_action_time -= (1 << G.poll_exp);
1587 VERB3 bb_error_msg("polladj: discipline_jitter:%f --poll_exp=%d",
1588 G.discipline_jitter, G.poll_exp);
1591 VERB3 bb_error_msg("polladj: decr:%d", G.polladj_count);
1596 /* Decide when to send new query for this peer */
1597 interval = poll_interval(0);
1599 set_next_and_close_sock:
1600 set_next(p, interval);
1601 /* We do not expect any more packets from this peer for now.
1602 * Closing the socket informs kernel about it.
1603 * We open a new socket when we send a new query.
1611 #if ENABLE_FEATURE_NTPD_SERVER
1612 static NOINLINE void
1613 recv_and_process_client_pkt(void /*int fd*/)
1617 len_and_sockaddr *to;
1618 struct sockaddr *from;
1620 uint8_t query_status;
1621 l_fixedpt_t query_xmttime;
1623 to = get_sock_lsa(G.listen_fd);
1624 from = xzalloc(to->len);
1626 size = recv_from_to(G.listen_fd, &msg, sizeof(msg), MSG_DONTWAIT, from, &to->u.sa, to->len);
1627 if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
1630 if (errno == EAGAIN)
1632 bb_perror_msg_and_die("recv");
1634 addr = xmalloc_sockaddr2dotted_noport(from);
1635 bb_error_msg("malformed packet received from %s: size %u", addr, (int)size);
1640 query_status = msg.m_status;
1641 query_xmttime = msg.m_xmttime;
1643 /* Build a reply packet */
1644 memset(&msg, 0, sizeof(msg));
1645 msg.m_status = G.stratum < MAXSTRAT ? G.ntp_status : LI_ALARM;
1646 msg.m_status |= (query_status & VERSION_MASK);
1647 msg.m_status |= ((query_status & MODE_MASK) == MODE_CLIENT) ?
1648 MODE_SERVER : MODE_SYM_PAS;
1649 msg.m_stratum = G.stratum;
1650 msg.m_ppoll = G.poll_exp;
1651 msg.m_precision_exp = G_precision_exp;
1652 /* this time was obtained between poll() and recv() */
1653 msg.m_rectime = d_to_lfp(G.cur_time);
1654 msg.m_xmttime = d_to_lfp(gettime1900d()); /* this instant */
1655 msg.m_reftime = d_to_lfp(G.reftime);
1656 msg.m_orgtime = query_xmttime;
1657 msg.m_rootdelay = d_to_sfp(G.rootdelay);
1658 //simple code does not do this, fix simple code!
1659 msg.m_rootdisp = d_to_sfp(G.rootdisp);
1660 version = (query_status & VERSION_MASK); /* ... >> VERSION_SHIFT - done below instead */
1661 msg.m_refid = G.refid; // (version > (3 << VERSION_SHIFT)) ? G.refid : G.refid3;
1663 /* We reply from the local address packet was sent to,
1664 * this makes to/from look swapped here: */
1665 do_sendto(G.listen_fd,
1666 /*from:*/ &to->u.sa, /*to:*/ from, /*addrlen:*/ to->len,
1675 /* Upstream ntpd's options:
1677 * -4 Force DNS resolution of host names to the IPv4 namespace.
1678 * -6 Force DNS resolution of host names to the IPv6 namespace.
1679 * -a Require cryptographic authentication for broadcast client,
1680 * multicast client and symmetric passive associations.
1681 * This is the default.
1682 * -A Do not require cryptographic authentication for broadcast client,
1683 * multicast client and symmetric passive associations.
1684 * This is almost never a good idea.
1685 * -b Enable the client to synchronize to broadcast servers.
1687 * Specify the name and path of the configuration file,
1688 * default /etc/ntp.conf
1689 * -d Specify debugging mode. This option may occur more than once,
1690 * with each occurrence indicating greater detail of display.
1692 * Specify debugging level directly.
1694 * Specify the name and path of the frequency file.
1695 * This is the same operation as the "driftfile FILE"
1696 * configuration command.
1697 * -g Normally, ntpd exits with a message to the system log
1698 * if the offset exceeds the panic threshold, which is 1000 s
1699 * by default. This option allows the time to be set to any value
1700 * without restriction; however, this can happen only once.
1701 * If the threshold is exceeded after that, ntpd will exit
1702 * with a message to the system log. This option can be used
1703 * with the -q and -x options. See the tinker command for other options.
1705 * Chroot the server to the directory jaildir. This option also implies
1706 * that the server attempts to drop root privileges at startup
1707 * (otherwise, chroot gives very little additional security).
1708 * You may need to also specify a -u option.
1710 * Specify the name and path of the symmetric key file,
1711 * default /etc/ntp/keys. This is the same operation
1712 * as the "keys FILE" configuration command.
1714 * Specify the name and path of the log file. The default
1715 * is the system log file. This is the same operation as
1716 * the "logfile FILE" configuration command.
1717 * -L Do not listen to virtual IPs. The default is to listen.
1719 * -N To the extent permitted by the operating system,
1720 * run the ntpd at the highest priority.
1722 * Specify the name and path of the file used to record the ntpd
1723 * process ID. This is the same operation as the "pidfile FILE"
1724 * configuration command.
1726 * To the extent permitted by the operating system,
1727 * run the ntpd at the specified priority.
1728 * -q Exit the ntpd just after the first time the clock is set.
1729 * This behavior mimics that of the ntpdate program, which is
1730 * to be retired. The -g and -x options can be used with this option.
1731 * Note: The kernel time discipline is disabled with this option.
1733 * Specify the default propagation delay from the broadcast/multicast
1734 * server to this client. This is necessary only if the delay
1735 * cannot be computed automatically by the protocol.
1737 * Specify the directory path for files created by the statistics
1738 * facility. This is the same operation as the "statsdir DIR"
1739 * configuration command.
1741 * Add a key number to the trusted key list. This option can occur
1744 * Specify a user, and optionally a group, to switch to.
1747 * Add a system variable listed by default.
1748 * -x Normally, the time is slewed if the offset is less than the step
1749 * threshold, which is 128 ms by default, and stepped if above
1750 * the threshold. This option sets the threshold to 600 s, which is
1751 * well within the accuracy window to set the clock manually.
1752 * Note: since the slew rate of typical Unix kernels is limited
1753 * to 0.5 ms/s, each second of adjustment requires an amortization
1754 * interval of 2000 s. Thus, an adjustment as much as 600 s
1755 * will take almost 14 days to complete. This option can be used
1756 * with the -g and -q options. See the tinker command for other options.
1757 * Note: The kernel time discipline is disabled with this option.
1760 /* By doing init in a separate function we decrease stack usage
1763 static NOINLINE void ntp_init(char **argv)
1771 bb_error_msg_and_die(bb_msg_you_must_be_root);
1773 /* Set some globals */
1774 G.stratum = MAXSTRAT;
1776 G.poll_exp = BURSTPOLL; /* speeds up initial sync */
1777 G.last_script_run = G.reftime = G.last_update_recv_time = gettime1900d(); /* sets G.cur_time too */
1781 opt_complementary = "dd:p::wn"; /* d: counter; p: list; -w implies -n */
1782 opts = getopt32(argv,
1784 "wp:S:"IF_FEATURE_NTPD_SERVER("l") /* NOT compat */
1786 "46aAbgL", /* compat, ignored */
1787 &peers, &G.script_name, &G.verbose);
1788 if (!(opts & (OPT_p|OPT_l)))
1790 // if (opts & OPT_x) /* disable stepping, only slew is allowed */
1791 // G.time_was_stepped = 1;
1793 add_peers(llist_pop(&peers));
1794 if (!(opts & OPT_n)) {
1795 bb_daemonize_or_rexec(DAEMON_DEVNULL_STDIO, argv);
1796 logmode = LOGMODE_NONE;
1798 #if ENABLE_FEATURE_NTPD_SERVER
1801 G.listen_fd = create_and_bind_dgram_or_die(NULL, 123);
1802 socket_want_pktinfo(G.listen_fd);
1803 setsockopt(G.listen_fd, IPPROTO_IP, IP_TOS, &const_IPTOS_LOWDELAY, sizeof(const_IPTOS_LOWDELAY));
1806 /* I hesitate to set -20 prio. -15 should be high enough for timekeeping */
1808 setpriority(PRIO_PROCESS, 0, -15);
1810 bb_signals((1 << SIGTERM) | (1 << SIGINT), record_signo);
1811 /* Removed SIGHUP here: */
1812 bb_signals((1 << SIGPIPE) | (1 << SIGCHLD), SIG_IGN);
1815 int ntpd_main(int argc UNUSED_PARAM, char **argv) MAIN_EXTERNALLY_VISIBLE;
1816 int ntpd_main(int argc UNUSED_PARAM, char **argv)
1824 memset(&G, 0, sizeof(G));
1825 SET_PTR_TO_GLOBALS(&G);
1829 /* If ENABLE_FEATURE_NTPD_SERVER, + 1 for listen_fd: */
1830 cnt = G.peer_cnt + ENABLE_FEATURE_NTPD_SERVER;
1831 idx2peer = xzalloc(sizeof(idx2peer[0]) * cnt);
1832 pfd = xzalloc(sizeof(pfd[0]) * cnt);
1834 /* Countdown: we never sync before we sent 5 packets to each peer
1835 * NB: if some peer is not responding, we may end up sending
1836 * fewer packets to it and more to other peers.
1837 * NB2: sync usually happens using 5-1=4 packets, since last reply
1838 * does not come back instantaneously.
1840 cnt = G.peer_cnt * 5;
1842 while (!bb_got_signal) {
1848 /* Nothing between here and poll() blocks for any significant time */
1850 nextaction = G.cur_time + 3600;
1853 #if ENABLE_FEATURE_NTPD_SERVER
1854 if (G.listen_fd != -1) {
1855 pfd[0].fd = G.listen_fd;
1856 pfd[0].events = POLLIN;
1860 /* Pass over peer list, send requests, time out on receives */
1861 for (item = G.ntp_peers; item != NULL; item = item->link) {
1862 peer_t *p = (peer_t *) item->data;
1864 if (p->next_action_time <= G.cur_time) {
1865 if (p->p_fd == -1) {
1866 /* Time to send new req */
1868 G.initial_poll_complete = 1;
1870 send_query_to_peer(p);
1872 /* Timed out waiting for reply */
1875 timeout = poll_interval(-2); /* -2: try a bit sooner */
1876 bb_error_msg("timed out waiting for %s, reach 0x%02x, next query in %us",
1877 p->p_dotted, p->reachable_bits, timeout);
1878 set_next(p, timeout);
1882 if (p->next_action_time < nextaction)
1883 nextaction = p->next_action_time;
1886 /* Wait for reply from this peer */
1887 pfd[i].fd = p->p_fd;
1888 pfd[i].events = POLLIN;
1894 timeout = nextaction - G.cur_time;
1897 timeout++; /* (nextaction - G.cur_time) rounds down, compensating */
1899 /* Here we may block */
1900 VERB2 bb_error_msg("poll %us, sockets:%u", timeout, i);
1901 nfds = poll(pfd, i, timeout * 1000);
1902 gettime1900d(); /* sets G.cur_time */
1904 if (G.adjtimex_was_done
1905 && G.cur_time - G.last_script_run > 11*60
1907 /* Useful for updating battery-backed RTC and such */
1908 run_script("periodic");
1909 gettime1900d(); /* sets G.cur_time */
1914 /* Process any received packets */
1916 #if ENABLE_FEATURE_NTPD_SERVER
1917 if (G.listen_fd != -1) {
1918 if (pfd[0].revents /* & (POLLIN|POLLERR)*/) {
1920 recv_and_process_client_pkt(/*G.listen_fd*/);
1921 gettime1900d(); /* sets G.cur_time */
1926 for (; nfds != 0 && j < i; j++) {
1927 if (pfd[j].revents /* & (POLLIN|POLLERR)*/) {
1929 recv_and_process_peer_pkt(idx2peer[j]);
1930 gettime1900d(); /* sets G.cur_time */
1933 } /* while (!bb_got_signal) */
1935 kill_myself_with_sig(bb_got_signal);
1943 /*** openntpd-4.6 uses only adjtime, not adjtimex ***/
1945 /*** ntp-4.2.6/ntpd/ntp_loopfilter.c - adjtimex usage ***/
1949 direct_freq(double fp_offset)
1954 * If the kernel is enabled, we need the residual offset to
1955 * calculate the frequency correction.
1957 if (pll_control && kern_enable) {
1958 memset(&ntv, 0, sizeof(ntv));
1961 clock_offset = ntv.offset / 1e9;
1962 #else /* STA_NANO */
1963 clock_offset = ntv.offset / 1e6;
1964 #endif /* STA_NANO */
1965 drift_comp = FREQTOD(ntv.freq);
1967 #endif /* KERNEL_PLL */
1968 set_freq((fp_offset - clock_offset) / (current_time - clock_epoch) + drift_comp);
1974 set_freq(double freq) /* frequency update */
1982 * If the kernel is enabled, update the kernel frequency.
1984 if (pll_control && kern_enable) {
1985 memset(&ntv, 0, sizeof(ntv));
1986 ntv.modes = MOD_FREQUENCY;
1987 ntv.freq = DTOFREQ(drift_comp);
1989 snprintf(tbuf, sizeof(tbuf), "kernel %.3f PPM", drift_comp * 1e6);
1990 report_event(EVNT_FSET, NULL, tbuf);
1992 snprintf(tbuf, sizeof(tbuf), "ntpd %.3f PPM", drift_comp * 1e6);
1993 report_event(EVNT_FSET, NULL, tbuf);
1995 #else /* KERNEL_PLL */
1996 snprintf(tbuf, sizeof(tbuf), "ntpd %.3f PPM", drift_comp * 1e6);
1997 report_event(EVNT_FSET, NULL, tbuf);
1998 #endif /* KERNEL_PLL */
2007 * This code segment works when clock adjustments are made using
2008 * precision time kernel support and the ntp_adjtime() system
2009 * call. This support is available in Solaris 2.6 and later,
2010 * Digital Unix 4.0 and later, FreeBSD, Linux and specially
2011 * modified kernels for HP-UX 9 and Ultrix 4. In the case of the
2012 * DECstation 5000/240 and Alpha AXP, additional kernel
2013 * modifications provide a true microsecond clock and nanosecond
2014 * clock, respectively.
2016 * Important note: The kernel discipline is used only if the
2017 * step threshold is less than 0.5 s, as anything higher can
2018 * lead to overflow problems. This might occur if some misguided
2019 * lad set the step threshold to something ridiculous.
2021 if (pll_control && kern_enable) {
2023 #define MOD_BITS (MOD_OFFSET | MOD_MAXERROR | MOD_ESTERROR | MOD_STATUS | MOD_TIMECONST)
2026 * We initialize the structure for the ntp_adjtime()
2027 * system call. We have to convert everything to
2028 * microseconds or nanoseconds first. Do not update the
2029 * system variables if the ext_enable flag is set. In
2030 * this case, the external clock driver will update the
2031 * variables, which will be read later by the local
2032 * clock driver. Afterwards, remember the time and
2033 * frequency offsets for jitter and stability values and
2034 * to update the frequency file.
2036 memset(&ntv, 0, sizeof(ntv));
2038 ntv.modes = MOD_STATUS;
2041 ntv.modes = MOD_BITS | MOD_NANO;
2042 #else /* STA_NANO */
2043 ntv.modes = MOD_BITS;
2044 #endif /* STA_NANO */
2045 if (clock_offset < 0)
2050 ntv.offset = (int32)(clock_offset * 1e9 + dtemp);
2051 ntv.constant = sys_poll;
2052 #else /* STA_NANO */
2053 ntv.offset = (int32)(clock_offset * 1e6 + dtemp);
2054 ntv.constant = sys_poll - 4;
2055 #endif /* STA_NANO */
2056 ntv.esterror = (u_int32)(clock_jitter * 1e6);
2057 ntv.maxerror = (u_int32)((sys_rootdelay / 2 + sys_rootdisp) * 1e6);
2058 ntv.status = STA_PLL;
2061 * Enable/disable the PPS if requested.
2064 if (!(pll_status & STA_PPSTIME))
2065 report_event(EVNT_KERN,
2066 NULL, "PPS enabled");
2067 ntv.status |= STA_PPSTIME | STA_PPSFREQ;
2069 if (pll_status & STA_PPSTIME)
2070 report_event(EVNT_KERN,
2071 NULL, "PPS disabled");
2072 ntv.status &= ~(STA_PPSTIME |
2075 if (sys_leap == LEAP_ADDSECOND)
2076 ntv.status |= STA_INS;
2077 else if (sys_leap == LEAP_DELSECOND)
2078 ntv.status |= STA_DEL;
2082 * Pass the stuff to the kernel. If it squeals, turn off
2083 * the pps. In any case, fetch the kernel offset,
2084 * frequency and jitter.
2086 if (ntp_adjtime(&ntv) == TIME_ERROR) {
2087 if (!(ntv.status & STA_PPSSIGNAL))
2088 report_event(EVNT_KERN, NULL,
2091 pll_status = ntv.status;
2093 clock_offset = ntv.offset / 1e9;
2094 #else /* STA_NANO */
2095 clock_offset = ntv.offset / 1e6;
2096 #endif /* STA_NANO */
2097 clock_frequency = FREQTOD(ntv.freq);
2100 * If the kernel PPS is lit, monitor its performance.
2102 if (ntv.status & STA_PPSTIME) {
2104 clock_jitter = ntv.jitter / 1e9;
2105 #else /* STA_NANO */
2106 clock_jitter = ntv.jitter / 1e6;
2107 #endif /* STA_NANO */
2110 #if defined(STA_NANO) && NTP_API == 4
2112 * If the TAI changes, update the kernel TAI.
2114 if (loop_tai != sys_tai) {
2116 ntv.modes = MOD_TAI;
2117 ntv.constant = sys_tai;
2120 #endif /* STA_NANO */
2122 #endif /* KERNEL_PLL */